Compares residual scoring against reference paths under windshield and yaw perturbations.
Public scorecard JSON enables third-party reproduction of charts.
Research & benchmarks
Inspect the evidence behind the algorithms.
Controlled figures, benchmark methodology, model boundaries, and reproducible product visuals for technical evaluation, not customer field-performance claims.
EvidenceControlledClearly labeled
FiguresReproducibleSource and model lineage
TiersPulse → ApexRuntime through forensic depth
BoundaryNo field claimUntil deployment-specific validation
01 / INTERACTIVE DIAGRAM
Benchmark snapshot
Controlled detection metrics — not field performance.
NADIR VISUAL · SEED DATAILLUSTRATIVE · NOT FIELD PERFORMANCE
Nominal Caution Critical
Illustrative seed-data visualization. Validate on your representative fleet sample before operational decisions.
02 / DETAILS
Two ways to read this page
Plain language for fleet and investor conversations. Technical detail for engineering diligence.
Plain-language summary
NADIR helps fleets detect likely ADAS miscalibration after repair using data you already export.
Shadow mode means advisory tiers and signed records — not automatic vehicle changes.
- Start with a four-week $750 proof pilot.
- Success = MTTFF-RE within seven days on most repair events.
- Expand only after a written pass/fail readout.
Technical summary
Batch ingest with schema validation, Pulse residual scoring, repair-event join, EVB export.
All outputs are versioned: model_id, policy_id, confidence, reason_codes.
- Mode
- Shadow advisory
- Actuation
- None
- Integrity
- HMAC-SHA256 EVB
03 / EVIDENCE STANDARD
Controlled benchmarks, not customer logos.
Figures on this site illustrate detector behavior on seed and simulation data.
Mobileye overlay study compares residual scoring against reference paths under documented scenarios.
Do not extrapolate figures to field accuracy without your own pilot readout.
- BenchmarkMobileye overlay scorecard
- FiguresDetection survival · FPR strata
- ReproPublic paper + JSON scorecard
04 / RESEARCH
Controlled benchmarks and reproducibility
Figures illustrate methodology — not field performance guarantees.
Useful for comparing policy thresholds — not for quoting customer-specific detection rates.
Glass procedures differ from front-end structural repairs — strata prevent averaged misleading metrics.
Canvas modules on marketing pages use seed fixtures — labeled illustrative throughout.
Academic-style methods sections for research; MTTFF-RE readouts for customers.
05 / DETECTION SCIENCE
Cross-modal residual scoring
How Pulse turns telematics into tiers after repair.
A green ADAS dashboard means self-check passed — not that alignment matches factory spec.
When camera lane proxy and radar range-rate disagree for days after glass work, that pattern elevates tier.
Single spikes from weather or harsh braking decay; persistence drives CAUTION and CRITICAL.
Pulse compares observed trajectories to a baseline envelope tuned for post-repair windows.
Outputs: tier, confidence, reason_codes[], model_id, policy_id.
Policies version independently from models — replay uses both IDs.
- Tiers
- NOMINAL · CAUTION · CRITICAL
- Policy
- Versioned JSON
- Outputs
- Reason codes + confidence
Sample CAUTION events with maintenance leads — label true vs nuisance.
Adjust observation window or procedure filters before week four readout.
Goal is actionable CRITICAL rate, not zero CAUTION.
Scoring window opens at RO close — pre-repair baseline excluded from post-RO residual.
Changepoint detectors flag abrupt shifts; sustained elevation drives tier transitions.
Cannot distinguish all root causes (wrong part vs skipped scan vs partial cal).
CRITICAL means inspect — not auto-assign blame to shop or OEM.
06 / GOVERNANCE
Shadow mode and product boundaries
Every public NADIR deployment starts advisory. These boundaries protect fleet legal teams and set honest expectations for insurers.
Shadow mode means NADIR scores telematics, assigns NOMINAL / CAUTION / CRITICAL tiers, and exports signed evidence — without writing calibration parameters to the vehicle.
Your maintenance team decides whether to pull a unit, schedule a bay recalibration, or return it to route. NADIR does not override OEM safety controllers.
This is intentional: post-repair detection pilots should prove alert quality before anyone discusses automation or correction paths.
- Console CRITICAL rows are recommendations, not work orders.
- Weekly digest emails summarize CAUTION trends without paging on-call at 2 a.m.
- Sample EVBs document what the model saw — useful for QA, not a legal verdict.
- Actuation
- None
- Default mode
- Shadow
- ECU writes
- Not supported in pilots
NADIR is not ASIL-certified software and does not replace OEM calibration procedures or static target rigs.
A valid evidence bundle signature proves integrity of the JSON record — not that the windshield camera is physically aligned.
Controlled benchmark figures on this site illustrate detector behavior on seed data; field accuracy requires your own MTTFF-RE readout.
If a vendor implies OEM sign-off from a dashboard alone, ask for the underlying evidence schema and pilot KPI.
After MTTFF-RE meets your SOW threshold, fleets may renew Shadow Ops ($750/mo) or open a regional evaluation ($15k+).
Correction, HORIZON forecasting, and tier upgrades remain internal or design-partner only until three paid shadow pilots complete.
Each expansion adds explicit contract language for data retention, rate limits, and excluded actuation paths.
- Pilot exit
- Pass / fail / stop
- Renewal gate
- MTTFF-RE
- Correction
- Not sold publicly
NADIR provides advisory analytics on exported telematics correlated to repair order close timestamps.
Outputs include tier assignments, reason codes, and optional signed JSON evidence bundles.
Customer retains operational authority over vehicle dispatch and calibration vendor selection.
SOW boundary snippet
# SOW boundary snippet (illustrative)
scope:
mode: shadow_advisory
ecu_writes: false
kpi: MTTFF-RE
actuation: none
deliverables:
- weekly_digest_csv
- console_critical_queue
- sample_evidence_bundles
- week_4_exhibit_a_chartThe commercial motion is a bounded $750 proof on 50–150 VINs with one KPI — easy for fleet CFOs to approve.
Shadow detection de-risks integration: no ECU access, no bay hardware, no ASIL claims.
Expansion revenue attaches to proven alert quality, not tier shopping in week one.
07 / PILOT KPI
MTTFF-RE — the one pilot metric
Mean Time To First Flag after Repair Event. One number for pass, fail, or stop.
When a vehicle leaves a glass or body shop, start a clock at repair order close.
MTTFF-RE measures hours until NADIR assigns CAUTION or CRITICAL in the post-repair observation window.
If most events flag within seven days, your maintenance team gets actionable lead time before silent misalignment becomes a safety conversation.
- Median MTTFF-RE ≤ 168 hours is the default pilot target.
- Pass rate ≥ 70% of in-cohort events is the companion statistic.
- Events without telemetry are excluded — never counted as passes.
Join repair_events.ro_closed_at to health_transitions where tier ∈ {CAUTION, CRITICAL}.
Delta = first_flag_at − ro_closed_at in UTC hours.
Appendix lists excluded VINs with reason: no_telemetry, invalid_vin, pre_existing_critical.
Exhibit A SQL (illustrative)
SELECT
re.vin,
re.ro_closed_at,
MIN(ht.scored_at) AS first_flag_at,
EXTRACT(EPOCH FROM (MIN(ht.scored_at) - re.ro_closed_at))/3600 AS mttff_hours
FROM repair_events re
JOIN health_transitions ht ON ht.vin = re.vin
AND ht.scored_at >= re.ro_closed_at
AND ht.tier IN ('CAUTION','CRITICAL')
GROUP BY 1,2Each dot or bar is one repair event in the pilot cohort.
Color encodes pass (≤168h) vs fail. Gray marks excluded rows.
Median line is the primary SLA statistic; mean is shown but not contractual.
- Primary stat
- Median hours
- Threshold
- 168h
- Coverage
- % with data
Cherry-picked VINs with perfect telematics while hard cases were dropped from cohort.
RO timestamps backfilled days after close — inflates MTTFF-RE artificially.
Treating missing data as NOMINAL instead of excluded.
NADIR readouts include an exclusion appendix specifically to prevent silent passes.
Define cohort size, date range, procedure types (glass, front-end, mirror), and telemetry coverage minimum.
State median and pass-rate thresholds explicitly. Define stop conditions if CAUTION noise exceeds agreed bounds.
Clarify that shadow mode continues regardless of pass — expansion is a commercial decision, not automatic actuation.
08 / REFERENCE TABLE
Reference matrix — research
Dense field reference for engineering diligence on this topic.
| Category | Field / control | Value | Notes |
|---|---|---|---|
| Pilot | mode | shadow_advisory | No ECU writes |
| Pilot | kpi | MTTFF-RE | Median ≤168h target |
| Pilot | cohort | 50–150 VINs | Shadow Proof SKU |
| Ingest | telemetry | POST /v1/telemetry/batch | CSV/JSONL |
| Ingest | repair | POST /v1/repair-events | RO anchor |
| Score | tier | NOMINAL|CAUTION|CRITICAL | Policy-bound |
| Score | model | pulse_residual_* | Versioned |
| Product | queue | Console CRITICAL rows | Operator-facing |
| Product | digest | Weekly CSV/email | CAUTION summary |
| Evidence | signature | HMAC-SHA256 | Integrity only |
| Evidence | verify | POST /v1/evidence/verify | Recompute digest |
| Security | transit | TLS 1.2+ | API only |
| Security | auth | Bearer + API key | Tenant scoped |
| Boundary | asil | Not claimed | Advisory analytics |
| Boundary | actuation | None | Shadow default |
| Commercial | proof | $750 flat | 4 weeks |
| Commercial | ops | $750/mo | After pass |
| Commercial | regional | $15k+ | 150–500 VINs |
10 / DETAILS
Common questions
Straight answers for buyers and reviewers.
Common questions
Straight answers for buyers and reviewers.
Regional fleets and repair networks proving post-repair ADAS detection first.
VIN, timestamps, telematics signals, and repair close times.
No. Shadow advisory only — detection, queue, and evidence export.
No. NADIR is advisory analytics, not safety-certified software.
11 / CONTROLLED FIGURES
Six views into detector behavior.
Each figure is labeled as controlled research or simulation output and links to the relevant technical explanation.
OPEN MODULEResidual traceNormalized drift statistic across a controlled sequence.Explore →
OPEN MODULEDetection survivalTime-to-detection behavior across model choices.Explore →
OPEN MODULEModality correlationCamera, radar, and proxy-signal relationships.Explore →
OPEN MODULEConformal FPRFalse-positive behavior by uncertainty stratum.Explore →
OPEN MODULEAlert heatmapFrames and scenarios that trigger detector response.Explore →
OPEN MODULETier comparisonPulse, Meridian, Apex, and Echo output depth.Explore →
12 / METHODOLOGY
A benchmark is only useful when lineage is complete.
NADIR separates source scenario, detector configuration, threshold policy, metric computation, and final interpretation.
NADIR / LAYERSSENSOR
RELIABILITY
SYSTEMCOMPOSABLE · VERSIONED · TRACEABLE
RELIABILITY
SYSTEMCOMPOSABLE · VERSIONED · TRACEABLE
01
SOURCE
Scenario
Dataset, synthetic perturbation, or Calibration Lab event.
02
CONFIG
Manifest
Sensor expectations, segment, tier, and model version.
03
EXECUTE
Run
Residuals, uncertainty, gates, and alerts.
04
MEASURE
Metrics
Detection time, recall, false positives, and stability.
05
REPORT
Evidence
Artifacts, assumptions, limitations, and reproducible IDs.
13 / EVIDENCE LEVELS
Keep controlled benchmarks separate from field claims.
The evaluation surface explicitly labels what each evidence type can support.
CAPABILITY
Controlled replayCalibration LabField deployment
Best for
RepeatabilityInteractive causalityOperational validity
Ground truth
Dataset / injectedScenario-definedOften partial
Environment
HistoricalSimulated browser labReal operations
Claim strength
Algorithm comparisonMechanism demonstrationDeployment-specific outcome
Current public use
AvailableAvailableDesign-partner stage
14 / EVALUATION WORKFLOW
Review a result without losing the assumptions.
A technical evaluation should be able to move backward from the chart to the exact source and configuration.
FIGURE
Read
Identify metric, units, sample, and comparison.
LINEAGE
Trace
Open scenario, manifest, model, and threshold.
RUN
Reproduce
Run replay or Calibration Lab configuration.
REVIEW
Challenge
Inspect limitations and alternative explanations.
NEXT
Decide
Choose whether field validation is warranted.
Evaluate NADIR on a defined scenario.
Bring a dataset, sensor path, or operating question and preserve every assumption.